iFlytek Aims to Match GPT-4 Turbo’s Performance with Upgrades to Spark Model in 2024

During the 2024 annual company gathering of iFlytek, Chairman Liu Qingfeng announced plans for significant upgrades to their Spark large model. The goal is to reach the current performance level of OpenAI’s GPT4-Turbo by the first half of the year.

The rise of OpenAI’s ChatGPT has spurred a wave of development in the field of large language models, with numerous domestic tech companies and research institutions joining the race in China. According to Liu, the year 2024 is expected to see a shift from a widespread competition involving hundreds of models to a more focused competition where a few key players emerge.

Since its initial release on May 6, 2023, iFlytek’s Spark model has been continuously upgraded, with the current version being V3.5. Reportedly, Spark V3.5 is close to the performance level of GPT-4 Turbo, even surpassing it in areas such as language comprehension and mathematical abilities. The model’s coding ability is said to be at 96% of GPT-4 Turbo’s level, and its multimodal understanding at 91% of that of GPT-4V.

Liu emphasized the importance of acknowledging and addressing the gap between Spark and GPT-4 Turbo in certain areas, such as complex knowledge reasoning, to ensure continued progress and growth. He also stressed the need for iFlytek to strive for leadership in specific areas and key industries while catching up with general large models.

However, to truly surpass ChatGPT, large models like Spark require extensive data and significant computational resources. In response to this, iFlytek has partnered with Huawei to develop an independently controllable computational base for large models, named “Feixing No.1”, leveraging Huawei’s Ascend ecosystem.